Measures
The study questionnaire comprised detailed questions about age, gender, education, current living situation, and sources of income (work, social benefits, dealing, theft and sex work) and amount of income from each income source. In this study dealing and theft were defined as illegal activities. The questionnaire also included questions about alcohol; the amount and frequency of cannabis use; the frequency of cocaine, LSD and ecstasy use. In addition, the respondents were asked if they had used heroin and if they had, their mode of intake (by injections, inhalation or smoking) and the amount of heroin in their last injection (if they had injected). The respondents were also asked about the age of their first injection, injection frequency and what substance they most commonly injected: heroin, amphetamine, both or other substances.
Furthermore, the questionnaire comprised questions about prescription drugs (frequency, type of drug and amount). In 1997, methadone and buprenorphine were not available as prescription drugs in Norway. Instead prescription drugs such as pain medication, sedatives, hypnotics and antiepileptic drugs were available. The respondents were also asked if they mixed prescription drugs and heroin and if they did, how often and what quantity of prescription drugs were used. The questionnaire is described in more detail elsewhere [
45].
The National Cause of Death Registry provided the dates of death and the primary causes of death of the participants. Causes of death were categorized by Statistics Norway according to the international classification system (ICD-9 codes). We divided the primary causes of death into six categories. The first category was acute intoxications with three subcategories “due to use of opioids (F11.0)”, “due to use of sedatives or hypnotics (F13.0)” and “Accidental poisoning by and exposure to narcotics and psychodysleptics [hallucinogens], not elsewhere classified (X42.0)”. The other categories were dependence syndrome (F11.2, F19.2), suicide (X70.0, X71.9), acute infections (A39.8, A41.9), chronic infections (B18.2, B20.7, B24.0), and other causes (C49.6, J45.9, K70.3, R99.8, V48.6, W74.8, X59.9 X99.8, Y21.8).
The OST programme in Oslo provided intake and discharge data, and The Norwegian Correctional Services provided incarceration dates and release dates. The information from the OST programme and The Norwegian Correctional Services were used in sub-analyses in a smaller dataset.
Variables and data analyses
Data analyses were completed using Stata version 13.0 [
46]. Chi square tests were used for the assessment of differences in baseline characteristics between genders. The cut-off points for the dummy variables “age” and “length of injection career” were based upon Darke and colleagues paper from 2011 [
3]. The cut-off point for “Total monthly income” was the median value for the total sample which was 33,000 Norwegian Kroner (NKR). This was approximately 3,560 Great British Pounds (GBP). The cut-off for the total monthly amount of heroin was the median amount of heroin used by the total sample (12.9 grams).
Crude mortality rate (CMR) was calculated by summing the person years (PY) contributed by each participant, by gender and calendar year, then summing the number of deaths by the same groups and calculating a rate per 1000 PY. Indirect standardized mortality ratio (SMR) was calculated by dividing the observed deaths in the cohort by the expected deaths if the cohort had the same specific rates as the death rate in the standard population. The standard population was the general population in Norway between 1997 and 2010 based on age and gender specific rates [
40]. The SMR was calculated using the age groups 15–19, 20–24, 25–29, 30–34, 35–39, 40–44, 45–49 and 50–54 years. All rates and ratios were reported with 95% Confidence Intervals (CI).
Two types of survival analyses were conducted. In the first analyses, time-at-risk was the period between the date of baseline interview (interviews conducted March, June or September 1997) and December 31st 2010. A continuous time model could thus be used. Incomplete spells were right-censored. The proportionality assumption for gender was not satisfied and therefore a proportional hazard model such as a Cox regression model could not be used. An Accelerated Failure Time (AFT) parametric model was used instead.
AFT models use log (time-to-failure), rather than risk (hazard) of failure [
47]. The regression coefficient B
*
k in AFT models summarizes the proportional effect on survival time T to a unit change in the corresponding covariate [
47]. However, it is more common to present the exponentiated regression coefficients, which are called time ratios (TR) [
47]. Therefore, this paper reports TR for each covariate’s estimates. TR ranges between 0 and infinity and a coefficient above 1 implies longer duration of survival, while a coefficient below 1 implies shorter duration [
47]. For example, if TR for men in a mortality study is three, it means that men have three times longer survival then women. On the other hand if TR for men is 0.3, then men have 70% shorter survival then women.
The three AFT models “Log-Logistic”, “Log-Normal” and “Weibull AFT” were assessed [
47]. The Log-Logistic model was chosen based upon an assessment of the AIC criterion and Log-Likelihood estimates.
Unobserved heterogeneity (“frailty”) was controlled for by estimating the models using a Gamma specification [
47]. To check robustness, the same assessment was conducted with Log-Normal and Weibull AFT models.
The multivariate Log-Logistic model was theoretically based and we used known risk factors for increased mortality among drug users. These factors were age, sex work, length of injecting career, injection frequency, combination of heroin and prescription drugs in injections and alcohol use [
9,
12,
13,
17,
18,
27,
48‐
50].
Almost all participants (90%) had injected daily or almost daily and therefore there was not enough variation in frequency of use to include this variable in the regression analyses. Since we adjusted for the combination of prescription drugs in heroin injections, we could not use “Any use of prescription drugs” or “heroin use” as separate variables.
There were no reports of income from sex work from men, while 21 of the 44 women did. Therefore sex work was only included as an independent variable in Model 3, where females were analysed separately from males. In model 4, males were analysed separately, excluding sex work from the model. It was not possible to control for frailty in model 3 and 4, which was most likely due to a small sample size.
In the second survival analysis, time-at-risk was the period between 1.1.1998 and 31.12.2004. “Total years in OST” was used as the variable for substitution treatment. “Total years in prison” was used as the variable for incarceration. “Prison release” was included in the model as a time-dependent covariate. Since data on imprisonment and prison release was available from 1.1.1997, data was left censored (imprisonment dates before 1.1.1998 were omitted from the analyses). Incomplete spells from 31.12.2004 were right censored. Data had to be split into incarceration episodes. Due to the organisation of data, it was not possible to use the Log-Logistic model. However, in this limited time model the proportionality assumption was satisfied also for gender. The proportionality assumptions were tested using Schoenfeld residuals, and scaled Schoenfeld residuals [
51]. A Cox regression survival model could therefore be applied and hazard ratios (HR) and 95% CI were reported.
Four models were assessed. Model 1 comprised of “total years in OST”, “total years in prison” and “prison release”, to measure the effect on mortality without the other variables. “Prison release” was the risk up until three weeks after release. In model 2, the original variables were added to the analyses. In model 3, women were analysed separately and in model 4, men were analysed separately. It was not possible to assess for unshared frailty (unobserved heterogeneity) in Cox regression analyses [
52] and consequently this was not assessed.
The differences in baseline characteristics between the women who reported income from sex work and those who did not, were assessed in post-hoc analysis. The reason for the post-hoc analysis was that sex work significantly decreased survival time in the Log-Logistic analyses and we wanted to determine possible reasons for this. One hypothesis was that high income from sex work allowed these women to consume more drugs and thereby shortening their survival time. To explore this further, we compared the women who reported income from sex work to women who did not. Fisher’s Exact Test was used. For the continuous variables of “total monthly income”, “number of prescription drugs used yesterday”, “amount of heroin per injection”, “number of days used heroin in the past month” and “total amount of heroin used in the past month” a two-sample t-test with equal variances was used.
For all analyses, the significance level was set at 5% level, unless otherwise stated in the text.